Human body three-dimensional joint estimation frame based on depth network and positioning method thereof

Abstract

The invention discloses a human body three-dimensional joint estimation frame based on a depth network and a positioning method thereof. The frame comprises a two-dimensional posture sub-network, which is used for pre-training on a two-dimensional posture data set to extract two-dimensional posture features and transmit them to a two-dimensional-three-dimensional conversion module to generate precise two-dimensional prediction posture; the two-dimensional-three-dimensional conversion module configured to receive the two-dimensional posture features extracted by the two-dimensional posture sub-network and convert them into a three-dimensional posture feature space, and generate a three-dimensional posture rough estimation with consistent time sequence; a 3D-two-dimensional projection module, configured to convert the rough estimation of the intermediate three-dimensional pose estimated by the two-dimensional-three-dimensional conversion module to the two-dimensional space to generate two-dimensional projection posture, which corrects the estimated three-dimensional posture by optimizing the consistency between the two-dimensional projection posture and the two-dimensional predictionposture, and finally outputs accurate three-dimensional posture estimation with time-space consistency and two-dimensional three-dimensional geometrical consistency. The invention can improve the accuracy of the three-dimensional joint prediction positioning of the human body.

Description

technical field [0001] The present invention relates to the fields of human body three-dimensional pose estimation, computer vision and human-computer interaction, in particular to a human body three-dimensional joint point estimation framework based on a deep network and a positioning method for improving the positioning accuracy of human body three-dimensional joint points based on a self-supervised correction mechanism. Background technique [0002] Pose estimation is an important field of computer vision research. Its main task is to enable computers to automatically perceive and understand human behavior. Its applications include intelligent monitoring, patient monitoring, and some systems involving human-computer interaction. The goal of human 3D joint point positioning is to automatically infer the specific position of the human body in the real world from images containing people, and reconstruct the motion of the human body through this information, laying the founda...

AI Technical Summary

Problems solved by technology

[0007] (3) In order to preserve the constraint relationship between the bones and joints of the human body, an extremely complex model is designed, resulting in poor scalability and generalization capabilities of the model

[0008] Most of the existing 3D joint point positioning methods rely on the complex a priori conditions of artificial design and the constraints of human bone structure, but the parameter setting of this method depends on the internal parameters of the sensor, and it is difficult to transplant to other camera sensors

Images